Deep brain stimulation

Benabid’s group was performing thalamotomy — deliberately destroying a small volume of thalamus to stop tremor — and using stimulation intraoperatively to confirm they were in the right place. They noticed that high-frequency stimulation itself suppressed the tremor, reversibly. If stimulation does what the lesion does, and can be switched off, there is no need to make the lesion.

That is the whole clinical argument, and it is a good one: reversible, adjustable, and titratable, where a lesion is none of those.

From one lesion to a therapy

The first protocol hedged. Patients with tremor on both sides had a thalamotomy on the worse side and, on the other, an electrode wired to a stimulator under the skin. That stimulator was a device sold for pain, and its programmable rate stopped at 130 Hz although 200 Hz looked better — so the frequency of the first chronic implants was capped by the hardware on the shelf.

By 1991 the Grenoble group had implanted electrodes in 43 thalami in 32 patients. Tremor stopped completely for 27 of the 43 and improved markedly for 11 more, and side effects went away when the stimulation was turned down or off. The conclusion was the argument above, now with data: stimulation was preferable to thalamotomy, especially when both sides of the brain needed treating. The FDA approved unilateral thalamic stimulation for tremor in 1997.

Worth being clear that chronic stimulation of deep structures was not invented in Grenoble. As Hariz and colleagues have documented, it had been used since the early 1950s, first in psychiatry, and earlier groups had argued for “high” frequencies too, without always saying how high. What Grenoble established was the substitution: a reversible setting doing the job of a specific, established lesion.

A loop you can intervene in

Thalamic stimulation suppresses tremor, but in Parkinson’s disease it leaves slowness and rigidity alone. The target for those came from a circuit diagram.

In 1986 Alexander, DeLong and Strick described the basal ganglia as stations on parallel loops — cortex to basal ganglia to thalamus, and back to a particular part of the frontal lobe. Models published in 1989 and 1990, by Albin, Young and Penney and by DeLong, added the pathology: with dopamine depleted, the subthalamic nucleus becomes overactive and basal ganglia output rises — too much output, too little movement. In 1990 Bergman, Wichmann and DeLong lesioned the subthalamic nucleus in monkeys made parkinsonian with the toxin MPTP, and akinesia, rigidity and tremor all improved on the opposite side of the body.

Stimulation then followed the lesion, as it had in the thalamus. High-frequency stimulation of the subthalamic nucleus relieved rigidity and slowness in two MPTP-treated monkeys in Bordeaux in 1993, and in 1995 Limousin, Benabid and colleagues reported the first patients stimulated there on both sides: in three people, motor scores improved by 42–84%. In their 1998 series of 24 patients, scores off medication had improved by 60% after a year, with the dopaminergic drug dose halved. The FDA approved subthalamic or pallidal stimulation for advanced Parkinson’s disease in 2002, and Benabid and DeLong shared the 2014 Lasker–DeBakey clinical award.

That is what the electrode gave the other side of the border: the basal ganglia as a loop you can intervene in. The circuit model said where to act; stimulation made the act reversible, so it could be tried in people and undone if wrong. Stimulation has since been approved for other disorders: dystonia in 2003, at the same two targets; obsessive-compulsive disorder in 2009, at the anterior limb of the internal capsule (explicitly as an alternative to the capsulotomy lesion); and epilepsy in 2018, at the anterior thalamic nucleus.

The loop also answered back, which is the more interesting half. In 2003 Hashimoto and colleagues stimulated the subthalamic nucleus of parkinsonian monkeys at settings that relieved their symptoms, and the internal pallidum — the output stage — fired faster, locked to the stimulus: the opposite of what a model built on too much output predicts. And the electrodes turned out to be sensors. Recording through them in the days after surgery, Brown and colleagues found the subthalamic nucleus and pallidum dominated, off medication, by synchronised activity below 30 Hz, which levodopa suppressed (2001). DeLong’s own later account, written with Thomas Wichmann in 2016, is about pattern rather than level: stimulation “replaces the abnormal basal ganglia output with a more tolerable pattern”.

The uncomfortable part

Nobody fully agrees on the mechanism. High-frequency stimulation was initially assumed to inhibit the target — because it mimicked a lesion — but it drives axons near the electrode to fire. Current explanations involve disrupting pathological oscillatory patterns in basal-ganglia loops rather than silencing a nucleus, and the honest summary is that the therapy is decades ahead of the account of it.

The axonal part is now well supported. Experiments had produced an apparent contradiction: activity suppressed in the stimulated nucleus, yet increased input to the nuclei it projects to. In 2004 McIntyre, Grill and colleagues resolved it by coupling a finite-element model of the clinical electrode to a multicompartment cable model of a thalamic neuron. Above threshold, the pulses suppressed the cell body’s own firing while the axon fired at the stimulus frequency — and the rest of the brain hears the axon. An output driven at a high, regular rate also carries no information about its inputs, which Grill and colleagues called an “informational lesion”: one way an excitatory stimulus can still behave like a cut. Benabid’s own guess, in 1996, had been “the inhibition or jamming of a retroactive loop”.

This should be read alongside TMS, where the same gap exists. Both are effective, both are in clinical use, and in neither case does the mechanism story constrain the parameters much. Which is precisely why the parameters are worth being able to vary.

Where the engineering sits

An implanted stimulator has constraints a bench instrument does not:

• Charge balance. Net DC current into tissue causes electrode corrosion and tissue damage, so every pulse must be followed by charge recovery. This is not a refinement; it is what makes chronic stimulation survivable.
• Electrode impedance drifts. Encapsulation tissue forms around an implant over weeks, changing the load. A stimulator specifying current rather than voltage is preferred largely because of this. Clinical stimulators were long voltage-controlled, and clinical impedances run from about 500 to 1,500 Ω; in one model, going from 790 to 1,244 Ω at a typical 3 V, 90 µs, 130 Hz setting shrank the volume of tissue activated by about half.
• Energy. The battery is inside the patient. Duty cycle and amplitude are limited by that as well as by physiology. In one early series, stimulators whose batteries ran out had lasted a median of 45 months, the hardest-driven failing first — and each replacement is another operation.

Conventional systems stimulate open-loop at fixed parameters, which is where closed-loop control has long been the obvious opportunity — and why it needs stimulators whose timing and waveform can be commanded.

Closing the loop

The idea is to stimulate only when the brain needs it, using subthalamic beta activity (13–30 Hz) — which tracks how impaired the patient is, and which stimulation suppresses — as the feedback signal. The difficulty is electrical: a rhythm of a few microvolts has to be sensed on the same lead that is delivering pulses of a few volts, around a million times larger. Part of the answer is geometry. Record differentially between the two contacts either side of the one stimulating, and the stimulus arrives at both as a common-mode signal that the amplifier can reject — provided the two contacts’ impedances match.

Little, Brown and colleagues closed the loop this way in 2013, in eight patients, in the days between implanting the electrodes and implanting the stimulator. The beta band was filtered, rectified and smoothed over 400 ms; stimulation switched on 30–40 ms after it crossed a threshold and off when it fell back, ramping over 250 ms each way to avoid paraesthesia. Motor scores improved by 66% (50% on blinded video rating), roughly 30% better than continuous stimulation, with stimulation on 56% less of the time. It was a proof of principle — ten-minute blocks, one side of the body, three rating items.

Twelve years later it is a product. In February 2025 the FDA approved an adaptive mode for Medtronic’s sensing-enabled stimulator, which adjusts amplitude within limits the clinician sets, according to power in a patient-specific band between 8 and 30 Hz. In one mode the amplitude ramps up whenever that power is above a single threshold and down whenever it is below; in the other it holds steady while the power stays between two thresholds. In control terms, a slew-limited bang-bang controller and one with a dead band. The pivotal trial enrolled a main cohort of 68 people already stable on continuous stimulation and tested each mode at home for a month. The bar was a performance goal rather than superiority — at least half the participants losing less than two hours a day of good “on” time — and both modes cleared it, though against a threshold revised after the fact, so the FDA treated the results as descriptive.

The energy story is more sober than 2013’s. Stimulation energy fell by 13–15% rather than by half, a saving the FDA did not judge clinically significant, and its review projected battery life about 5% longer than continuous stimulation in the dual-threshold mode but 4% shorter in the single-threshold mode, partly because that algorithm samples at 20 Hz rather than 5 Hz. Sensing and deciding draw on the same battery as stimulating — continuous sensing alone cost about 3% of projected battery life — so a controller has to save more than it costs to run.

What it costs

It is brain surgery, and the risk is real. In a German randomised trial of 156 patients, stimulation beat medication alone on quality of life and motor scores, but serious adverse events were three times as common — 13% against 4% — and included a fatal intracerebral haemorrhage. The stimulation has side effects of its own, mainly neurocognitive, and others from current spreading into neighbouring structures, depending on exactly where the electrode sits.

In Parkinson’s it largely treats what levodopa treats. In the 1998 series, motor scores improved by 60% off medication but by only 10% on it, and the FDA indication is restricted to levodopa-responsive disease: stimulation mostly makes the bad hours better, not the good hours better still.

And two of the later indications cleared a lower bar. Dystonia and obsessive-compulsive disorder were approved as humanitarian device exemptions, a route for small patient populations that is exempt from the usual requirement to show effectiveness.